Vault PKM

Give Claude persistent, structured memory across conversations using your Obsidian vault. Read, write, search, and navigate your knowledge base — all from within Claude Code.

Under the hood, this Claude Code plugin provides 20 MCP (Model Context Protocol) tools for note creation, semantic search, graph traversal, metadata queries, and session memory — plus agents, hooks, and skills for seamless workflow integration. Published on npm as obsidian-pkm.

If you find this useful, please star the repo — it helps others discover the plugin.

Watch the demo video

Why

Claude Code has built-in memory, but it's flat text files scoped to individual projects — no structure, no search beyond exact matches, no connections between notes, and no way to query across projects. As knowledge grows, it doesn't scale. This server replaces that with a proper PKM layer: structured notes with enforced metadata, semantic search, a navigable knowledge graph, and cross-project access through a single Obsidian vault.

• Structured session memory — Every tool call is logged with timestamps and session IDs, so Claude can recall exactly what was read, written, and searched in previous conversations — not just what was saved to a text file.
• Structured knowledge creation — ADRs, research notes, devlogs, and tasks are created from enforced templates with validated frontmatter — not freeform text dumps. Your vault stays consistent and queryable.
• Semantic discovery — "Find my notes about caching strategies" works even if you never used the word "caching." Conceptual search surfaces relevant knowledge that keyword search misses.
• Graph-aware connections — Claude explores your knowledge graph by following wikilinks, discovering related notes by proximity rather than just content. Link suggestions help weave new notes into your existing web of knowledge.
• Knowledge capture — Decisions, tasks, and research findings are captured by specialized agents in the background without interrupting your coding flow.

Without this, knowledge stays fragmented across per-project memory files and chat logs. With it, your AI assistant maintains a unified knowledge base that compounds over time.

How It Compares

remember is great for lightweight session continuity with minimal setup. Vault PKM is for developers who want a structured, searchable, interconnected knowledge base that grows with every project.

Features

Knowledge Creation & Editing

Discovery & Search

Graph & Connections

Reading & Navigation

Organization & Maintenance

Session Memory

Agents, Skills & Commands

Agents (3) run autonomously in foreground or background:

Skills (6) are guided workflows triggered by slash commands:

Commands (2) for setup and configuration:

Prerequisites

• Node.js >= 20 (Node 18 is EOL; uses native fetch and ES modules)
• An MCP-compatible client such as Claude Code

Prebuilt native binaries are included for Node 20/22 on Linux x64, macOS (x64/arm64), and Windows x64. Most users need nothing else. If the prebuilt fails, you'll need C++ build tools — see Troubleshooting.

Quick Start

1. Install the Plugin

claude plugin marketplace add anthropics/claude-plugins-community
claude plugin install vault-pkm@claude-community

2. Configure

Run the setup skill in Claude Code:

/vault-pkm:setup

The setup skill walks you through vault path, API keys, tool permissions, and verification. Hooks are registered automatically by the plugin system.

Important: Restart your Claude Code session after setup completes so the MCP server picks up the new configuration.

3. Scaffold Your Vault (optional)

If you need templates and the PARA folder structure, run the vault scaffolding wizard:

npx obsidian-pkm init

This is separate from the plugin install above — it only sets up your vault's directory structure (PARA folders, note templates). Nothing is written until you confirm each step.

Note: The first npx run downloads and compiles native dependencies, which may take 30-60 seconds. Subsequent runs are instant.

4. Verify It Works

Open Claude Code and try:

List the folders in my vault

Claude should call vault_list and show your vault's directory structure. If it works, the server is connected and ready.

5. Enable Semantic Search (optional)

Add your OpenAI API key to ~/.claude/settings.json under the env block:

{
  "env": {
    "VAULT_PATH": "/path/to/vault",
    "VAULT_PKM_OPENAI_KEY": "sk-your-key-here"
  }
}

Use VAULT_PKM_OPENAI_KEY (preferred) to avoid conflicts with project-level OpenAI keys. OPENAI_API_KEY is also accepted as a fallback. The previously-documented OBSIDIAN_PKM_OPENAI_KEY still works as a deprecated fallback — plan to rename it in your config. Restart Claude Code after saving.

Optional: VAULT_PKM_VAULT_NAME — overrides the vault name used in obsidian:// links emitted in tool output. Defaults to basename(VAULT_PATH), which is correct for most setups. Set this only if your on-disk folder name differs from the vault name you registered in Obsidian (e.g. VAULT_PATH=/Users/me/Notes but Obsidian shows the vault as My PKM).

This enables vault_semantic_search and vault_suggest_links. Uses text-embedding-3-large with a SQLite + sqlite-vec index stored at .obsidian/semantic-index.db. The index rebuilds automatically — delete the DB file to force a full re-embed.

Vault/
├── 00-Inbox/
├── 01-Projects/
│   └── ProjectName/
│       ├── _index.md
│       ├── planning/
│       ├── research/
│       └── development/decisions/
├── 02-Areas/
├── 03-Resources/
├── 04-Archive/
├── 05-Templates/          # Note templates loaded by vault_write
└── 06-System/

graph LR
    CC[Claude Code] -->|MCP protocol| IDX[index.js]
    IDX --> HND[handlers.js]
    IDX --> E[embeddings.js]
    IDX --> A[activity.js]
    HND --> H[helpers.js]
    HND --> G[graph.js]
    HND --> U[utils.js]
    HND -->|read/write| V[(Obsidian Vault)]
    E -->|embeddings API| OAI[OpenAI]
    E -->|vector store| DB[(SQLite + sqlite-vec)]
    A -->|activity log| DB2[(SQLite)]

├── index.js          # MCP server setup, tool registration, lifecycle
├── handlers.js       # Tool handler implementations
├── helpers.js        # Pure functions (path security, filtering, templates, frontmatter)
├── graph.js          # Wikilink resolution and BFS graph traversal
├── embeddings.js     # Semantic index (OpenAI embeddings, SQLite + sqlite-vec)
├── activity.js       # Activity log (session tracking, SQLite)
├── utils.js          # Shared utilities (frontmatter parsing, file listing)
├── cli.js            # CLI entry point (routes `init` subcommand or starts server)
├── init.js           # Vault scaffolding wizard (templates, PARA folders)
├── .claude-plugin/   # Plugin packaging
│   └── plugin.json   # Plugin manifest (identity, components, permissions)
├── hooks/            # Claude Code hooks (context loading, project resolution, session start)
├── agents/           # Specialized agents (vault-explorer, pkm-capture, link-auditor)
├── skills/           # PKM workflow skills (pkm-write, pkm-explore, pkm-session-end, add-task, triage-tasks, tackle-task)
├── commands/         # Slash commands (setup, init-project)
├── templates/        # Obsidian note templates
├── tests/            # Test suite (Node.js built-in test runner)
├── sample-project/   # Sample CLAUDE.md for your repos
└── docs/             # Supplementary documentation

How It Works

Knowledge creation is template-based. vault_write loads templates from 05-Templates/, substitutes Templater-compatible variables (<% tp.date.now("YYYY-MM-DD") %>, <% tp.file.title %>), and validates required frontmatter fields (type, created, tags). This ensures every note in your vault has consistent metadata — making it queryable, sortable, and discoverable from day one. Task notes enforce enum validation on status and priority; other types accept project, deciders, due, and source.

Knowledge discovery works at two levels. Keyword search (vault_search) finds exact terms. Semantic search embeds notes using OpenAI and finds conceptually related content — so "managing overwhelm" surfaces notes about "cognitive load" even if those exact words never appear together. The semantic index watches for file changes in real-time and syncs across machines via Obsidian Sync.

Knowledge connections are maintained through Obsidian's [[wikilink]] graph. vault_neighborhood traverses links via BFS to discover related notes by proximity, while vault_suggest_links recommends connections you haven't made yet. vault_move rewrites wikilinks across the vault when you reorganize, and vault_trash warns about links that would break.

Session memory records every tool call with timestamps and session IDs, so Claude can recall what was read, written, or searched in previous conversations. This turns ephemeral chat sessions into a continuous thread of work.

Knowledge capture uses the pkm-capture agent to update the project devlog and persist PKM-worthy content from a session (decisions, research findings, tasks, bug root causes) in one pass. It is triggered automatically after git commits via a PreToolUse hook, or can be dispatched manually after significant work blocks. Runs in the background without interrupting the coding flow.

Fuzzy path resolution lets read-only tools accept short names instead of full vault paths. vault_read({ path: "devlog" }) resolves to 01-Projects/MyApp/development/devlog.md automatically (.md extension optional). Folder-scoped tools like vault_list, vault_search, and vault_query accept partial folder names — folder: "MyApp" resolves to 01-Projects/MyApp. Ambiguous matches return an error listing candidates. Write/destructive tools always require exact paths.

Troubleshooting

better-sqlite3 build fails during install You need C++ build tools. See Prerequisites for your platform. On Linux, sudo apt install build-essential python3 usually fixes it.

Server starts but all tool calls fail with ENOENT Your VAULT_PATH is wrong or missing. The server validates this at startup and exits with a clear error. Run /vault-pkm:setup to reconfigure the vault path.

vault_write says "no templates available" Run npx obsidian-pkm init to install templates, or copy the templates/ files from this repo into your vault's 05-Templates/ directory. The server loads templates from there at startup.

Semantic search not appearing in tool list Set VAULT_PKM_OPENAI_KEY in ~/.claude/settings.json. See Enable Semantic Search. Without it, vault_semantic_search and vault_suggest_links are hidden entirely.

Server not showing up in Claude Code after install Run claude mcp list to check. If vault-pkm is missing, reinstall the plugin: claude plugin install vault-pkm@claude-community. Then run /vault-pkm:setup to configure it.

Semantic index not updating after file changes Check your Node version with node -v. The file watcher uses fs.watch({ recursive: true }) which requires Node.js >= 20.

Not sure if everything is set up correctly? Run npx obsidian-pkm doctor for a diagnostic checklist that validates your Node version, vault path, templates, API keys, and native dependencies.

Contributing

Contributions are welcome! Please read CONTRIBUTING.md for development setup, code style guidelines, and the pull request process before submitting changes.

See CHANGELOG.md for release history, SECURITY.md to report vulnerabilities, and PRIVACY.md for how the plugin handles your data.

License

MIT